In August, a developer I had mentored through The Open Ledger sent me a two-minute screen recording from his apartment in Kasarani. He had built an agent that watched airfare and booked a ticket when the price dipped below a threshold. It worked eleven times out of fourteen. The three failures were not inference failures. One booking site expired his session mid-transaction; one payment confirmation timed out; and in the third case, nobody โ not the merchant, not him, not the agent โ could produce a record of who had actually consented to what.
I kept that recording open while reading through a four-month retrospective of Google's I/O announcements that reached me here in Nairobi this September. The event itself was in May. Google had moved Search from a request-response appliance to a standing population of persistent agents: background monitors watching pages, blogs and social feeds around the clock, generative interfaces assembled per query, a booking flow covering the entire United States, telephone agents that call merchants who have no API, an eighth-generation TPU underneath, and a capital expenditure figure of roughly $180 to $190 billion โ about six times what Alphabet spent in 2022.
I want to be careful with that number. The first version of the analysis I read carried it as $18 to $19 billion, off by an order of magnitude, which would have changed every conclusion downstream. The scale distinction matters: this is not incremental spending on a feature. It is the price of believing that inference demand has structurally changed โ that always-on agents produce a baseload of compute where search only ever produced pulses.
Everything else in the announcement was engineering, not architecture. Fast models at half the price. Generative UI is code generation plus sandboxed rendering. Antigravity was repositioned from an encoding environment into a platform for managing fleets of autonomous agents. Each is a known technique in a new arrangement. I spent six months in 2017 reviewing more than 150 token standard drafts for the ZEIP-20 working group, and I learned then that the interesting question about a specification is never what it claims but where the implementation quietly concentrates power. Forty-two edge cases in ERC-20 transfer logic favored centralized validators. Nobody wrote them on purpose. The spec said neutral; the code voted.
The same reading applies here, and it is the part crypto should be taking notes on. Google did not choose Flash as its default because it was the strongest model. It chose it because it was four times faster and less than half the cost. That is a confession that the model layer has commoditized โ and once capability is purchasable by anyone, the only remaining moat is distribution. Search, Chrome, Android: three defaults stacked on one user. The agent is not the product. The default is the product.
Crypto has spent two bull markets arguing about which chain has the better virtual machine while losing the distribution argument to wallets, exchanges and app stores. Building libraries where others build empires remains a choice, and often the right one โ but we should stop calling it a strategy for winning.
What Google's framing admits, carefully, is that the governance layer is missing. An agent that books, pays and calls on your behalf needs identity, scoped permission, revocation, and a durable record of responsibility. Crypto has nominally been building these primitives for a decade. I am not convinced they work. Upgrade rights on almost every "decentralized" authorization contract still sit with a handful of multi-signature holders, which makes revocation a support ticket dressed as a function call. A user cannot verify that a permission was truly withdrawn. Ethics is not a feature; it is the foundation โ and a foundation you can pause with three of five keys is not a foundation.
The fragility is familiar in another register. In DeFi the loss never happens in the deterministic core of a contract; it happens at the seam where the chain meets a world it cannot see. Oracle feeds and their latency are that same seam as a voice agent navigating a phone tree in a language it half-understands. In both cases the industry's answer has been to decentralize the last mile in name while operating it from a small set of machines sharing one upgrade path. Chainlink demonstrated that a network can be called decentralized and still concentrate the consequential decisions. The last mile is always centralized, and it is always where the money disappears.
There is a metric problem too, and it mirrors ours. The adoption figures โ a billion monthly users, two and a half billion reaching AI summaries โ are self-reported and almost certainly include people who never intended to use anything. A record high in query volume can just as easily mean users are rewriting queries to escape an AI summary. That is query inflation, not demand. We have a word for this on-chain: unique wallets. Listening to the silence between the blocks means noticing what is not counted.
Here is where I part company with the enthusiasm. Everyone is debating whether these agents will work. The harder question is whether anyone will be permitted to let them work. Europe's Digital Markets Act already forces search choice screens, and American antitrust remedies are still being drafted. If default agent behavior is classified as gatekeeping, the moat gets decoupled by law rather than by competitors. The AI industry treats governance as a technical backlog. It is not. It is an unfinished negotiation about liability โ who pays when the agent books the wrong flight โ which is insurance and contract law, precisely the territory crypto already occupies with escrow, staking and slashing. The AI world is ignoring the only serious liability machinery that already exists, and we are too busy speculating to notice the opening.
One thing the retrospective never mentions, and should: these agents will not arrive everywhere at once. Mini-program ecosystems make transactional APIs far more common in parts of Asia than the open web does, which means agent commerce may land faster where the open web is weakest. Agentification is not one global process. Watching a mentored developer's agent succeed eleven times out of fourteen, I think that divergence is where the real work sits.
The next two years will be decided by who holds the permission, who can revoke it, and who can prove it afterward. Everything else is a demo.